Transformative Insights: Anticipating Parkinson’s Motor Fluctuations with Wearable Sensors and Advanced Deep Learning Models

Abstract

Parkinson’s disease is a typical and very invalidating neurogenerative disorder characterized by specific motor symptoms: bradykinesia, tremor, and dyskinesias. All of these significantly decrease the quality of life of patients. Classic diagnostic methods fail to define the disease’s ‘true’ state of the disease. This work presents a novel technique that aims to enhance the detection and classification of Parkinson’s disease motor symptoms using transformer-based models and wearable sensors. In this work, accelerometers at the wrists and waist to capture movement patterns related to PD, allowing the model to identify slight motor abnormalities. The results obtained were based on accuracy, recall, precision, and F1 score, all of which were observed at different attention heads settings. The efficiency is proven to improve the diagnostic results in comparison with traditional methods of binary classification having 98% and multilabel classification at 97%. More importantly, for the given model, the accuracy as recorded was approximately 99 percent. 34% with 8 attention heads on the GeneActiv dataset was recorded. The metrics of both the Pebble and smartphone datasets were also fairly accurate and exhibited nearly equal numbers of true positives, true negatives, false positives, and false negatives thus emphasizing the significance of the location of sensors and the complexity of the models in making accurate diagnoses. This work demonstrates the applicability of the transformer models in dealing with data interactions involved in the PD symptoms and calls for state-of-the-art PD diagnostic models capable of giving dynamic, precise, and personalized disease monitoring and management. The combination of artificial intelligence with wearable devices is a major achievement that affirms the possibility of improving the lives of patients with Parkinson’s Disease greatly.